Version 2.2.1

Multiple Back-Propagation v. 2.2.1 is now available

This version corrects a bug that prevented data files in a format other than CSV to be properly read (As a result MBP would crash when training a network). When using CUDA the value of the max step size (configuration) is now taken into account (in previous versions the value of 10.0 was used regardless of the max step size specified).

Multiple Back-Propagation Version 2.2

Version 2.2 of Multiple Back-Propagation (MBP) supports training networks with data containing missing values (data must be on a CSV file). A value is considered as being absent (missing) if its value is not present or if it is equal to a question mark (?). To deal with missing values MBP does not use imputation or any other pre-processing technique, but it rather creates a network with “selective inputs”, which can handle directly missing values. Selective inputs account for the creation of different (conceptual) models, bounded to each other and sharing information, demonstrated to be more accurate than the ones that can be created by other techniques and algorithms. Moreover neural networks with selective inputs are prepared to deal with faulty sensors. More information can be found on the paper: Lopes, N. and Ribeiro, B. “A Strategy for Dealing with Missing Values by using Selective Activation Neurons in a Multi-Topology Framework” to be published on the IJCNN 2010 conference.